微调与托管推理

Now branded Gemini Enterprise Agent Platform. PaLM 2 and Codey tuning went away with those models. The Gemini 2.5 targets retire on Google Cloud on 2026-10-20, so new tunes should start on Gemini 3.5 Flash or 3.1 Flash-Lite. Also a first-party path for tuning open models such as Gemma 3, Llama, and Qwen3.

产品特性：Supervised fine-tuning with adapter sizes from 1 to 16；Preference tuning on Gemini 2.5 Flash and Flash-Lite；Reinforcement learning fine-tuning on Gemini 3.5 Flash；Tuning checkpoints and continuous tuning。

Mistral now files its fine-tuning API under deprecated features and says it is no longer actively supported. Jobs still carry the $4 minimum and $2 monthly storage fee. Fine for keeping an existing tune alive, not a place to start new work.

产品特性：Supervised fine-tuning for text and vision；Classifier Factory for custom classifiers；European data residency。

Nova is now the center of Bedrock customization, and on-demand serving for tuned Nova models removes the old Provisioned Throughput cost at low volume. The Claude path is gone: Claude 3 Haiku was the last customizable Claude model and reached Bedrock end of life on 2026-09-10.

产品特性：Supervised fine-tuning and distillation；Reinforcement fine-tuning with GRPO；Continued pre-training；Nova Forge subscription for deeper customization。

Clear size-tiered pricing and no serving markup on tuned models, so many variants can share one base cheaply. No longer LoRA-only: full-parameter training and RFT are both on the price list.

产品特性：LoRA and full-parameter SFT and DPO；Reinforcement fine-tuning at on-demand GPU rates；Multi-LoRA serving at base-model prices。

Hugging Face now marks AutoTrain as no longer maintained, with no new features or bug fixes, and recommends Axolotl, TRL, or transformers Trainer instead. NVIDIA agreed on 2026-09-03 to acquire Hugging Face for $12.93B and says the Hub will stay open and accelerator-neutral.

产品特性：No-code UI in a Docker Space；Local UI and Python API；Push to Hub。

BYO fine-tuning infra. You own the code; Modal owns the GPU scheduling. The right answer if your tuning recipe does not fit a turn-key provider.

产品特性：Full BYO control；Serverless GPU；Distributed training；Open-source recipes。

CoreWeave agreed to acquire OpenPipe in September 2025. OpenPipe said it would keep supporting existing customers, the ART library, and managed SFT, while warning the product may evolve. The old distill-from-GPT-4-traces pitch is now part of a larger CoreWeave training stack, so older published pricing should not be relied on.

产品特性：Agent Reinforcement Trainer (ART) open-source RL library；Managed SFT workflows；Sits alongside Weights & Biases inside CoreWeave。

Rubrik agreed to acquire Predibase in June 2025, and predibase.com now redirects to Rubrik Agent Cloud. The standalone self-serve fine-tuning product is no longer marketed on its own; LoRAX, the open-source multi-LoRA server, is the piece most teams can still use directly.

产品特性：Multi-adapter LoRA serving；LoRAX open-source serving project still maintained on GitHub。

Strongest image LoRA experience (FLUX, SDXL). Cloudflare agreed to acquire Replicate in November 2025 to fold its model catalog into Workers. Less competitive for text-LLM fine-tuning than Together or Fireworks.

产品特性：Image LoRA fine-tuning standout；API-driven；Fast-booting fine-tunes without idle charges。

A research-grade training API rather than a job runner: you write the training loop and Thinking Machines runs the GPUs. Checkpoints can be downloaded, and pricing follows active parameters, which makes trillion-parameter MoE models such as Kimi-K2.6 practical to tune.

产品特性：Low-level API: forward_backward, optim_step, sample, save_state；Built-in RL losses (PPO, CISPO, DRO) and custom losses；Downloadable LoRA checkpoints, stored at $0.10/GB-month；Priced on active parameters, which favors MoE models。

A broad hosted menu of open models to tune, from Qwen3.5 0.8B up to Kimi K2.6, with a published price and minimum charge for every one. The default for teams that want a custom open model without writing a training loop.

产品特性：LoRA and full fine-tuning；SFT and DPO；Published per-model prices and minimums；Frontier-scale open models including Kimi K2.6 and DeepSeek V4 Flash。

Effectively closed. Bedrock's fine-tuning table lists Claude 3 Haiku as the only customizable Claude model; it entered Legacy on 2026-03-10, which blocks new fine-tuning jobs, and reached end of life on 2026-09-10. No Claude 4 or Claude 5 model appears in the table.

产品特性：Bedrock-only (AWS)；Claude 3 Haiku was the only customizable Claude model；Legacy status blocks new fine-tuning jobs。

OpenAI is winding the platform down. New organizations lost access on 2026-05-07, organizations with no fine-tuned inference in the prior 60 days lost job creation on 2026-07-02, and every remaining customer stops creating jobs on 2027-01-06. Tunes built on gpt-4.1-nano and o4-mini stop serving when those bases shut down on 2026-10-23. No model export. Useful now only for maintaining existing tunes.

产品特性：Supervised, DPO, vision, and reinforcement fine-tuning；Closed to new organizations since 2026-05-07；Fine-tunes keep serving until their base model is deprecated。

先看流量规模：月 token 量小、请求稀疏时按 token 付费几乎总是更便宜，因为闲置的卡不会给你打折；月 token 量大且稳定时，租卡自建的边际成本会低于按 token 的单价。用这页的月成本表对比成本测算里同规模的租卡月成本即可——注意把自建的运维与弹性损失也考虑进去。

LoRA/QLoRA 训练成本低一个量级、可插拔多套适配器，适合风格对齐与领域问答；全参微调效果上限更高但成本高、且每次迭代都要重训全量权重，适合基座能力确实不足的场景。表中的「微调方法」列可以直接看出各家支持范围：全参、LoRA、DPO、RLHF（或 RFT）的支持差异很大。

OpenAI 已公布自助微调的下线时间表（分阶段关闭新建任务），Anthropic 的 Bedrock 微调也进入 Legacy 不再接受新任务。结论很直接：闭源模型厂商正在收缩微调能力，长期依赖单一厂商的微调服务有停服风险。要做长期投入，优先选可导出权重或基于开源模型的托管微调。

很重要且常被忽略。同样是每百万 token 报价，上下文窗口决定你能不能塞进长文档（超出直接报错），输出速度决定用户体验（tok/s 低到个位数时交互体验会很差）。表中两列都来自公开来源实测/标注，缺值的地方保持空白。